Spectroscopic characterisation of feedstock for copper smelters by machine-learning

ABSTRACT

A flame-emission spectrometer was built to determine the elemental composition of powdered minerals that are important in copper smelting processes. The feedstock, consisting of milled concentrate, was fed into an oxyacetylene flame without sample preparation. The elemental composition (Cu, Fe, S, Si, and Zn) was determined by applying an artificial neural network (ANN) to a set of emission spectra obtained from Cu and Fe pure elemental powders, five pure mineral powders of known composition and 30 binary mixtures of these mineral samples. The ANN model was able to accurately predict the Cu and Fe content of these mineral powders within better than 2% of the value obtained from ICP-OES. The analysis was repeated on 12 industrial samples with well-known compositions. Spectra from these samples were analyzed both in isolation of the reference minerals and together with the reference minerals, giving similar results.

KEYWORDS:

• Machine learning
• artificial neural network
• copper flash furnace
• feedstock analysis
• flame emission spectroscopy
• process control

Acknowledgements

The authors acknowledge financial support by NSERC through the Alliance programme, by the National Research Council through the Industrial Research Assistance Program, and by Kingston Process Metallurgy, KPM. This work is part of the ProCuPro initative in collaboration with the University of Potsdam, Germany. An (unnamed) industrial partner generously provided the concentrate samples. HPL thanks Daniel German for critical comments on the manuscript, AB thanks Russell Dawes and Mark Woodrow for technical support.

Disclosure statement

BD is the co-founder and co-owner of KPM, a company that supported this research.

Additional information

Funding

This work was supported by Kingston Process Metallurgyand NSERC of Canada.